Where does AI actually sit in a disability claim today?

Rarely at the final denial, and almost everywhere before it. The common deployments are intake triage that routes a claim to a fast-track or an investigative queue; duration prediction that compares a diagnosis and occupation against expected return-to-work windows; risk scoring that ranks claims for referral to a medical reviewer, a vocational assessment or surveillance; document analysis that extracts restrictions and limitations from medical records; and correspondence generation that drafts the request-for-information and denial letters. Carriers describe all of this as decision support because a human signs the outcome. The legal question is not who signs it — it is whether the model changed the outcome, and a triage system that decides which claims receive scrutiny changes outcomes systematically even when no model ever writes a denial.

What does ERISA's full-and-fair-review standard require when a model is involved?

An adverse benefit determination has to tell the claimant the specific reasons for the denial, the plan provisions relied on, and what additional material would perfect the claim, and the claimant is entitled on request to the documents, records and other information relevant to the claim — including internal rules, guidelines, protocols or similar criteria relied upon. A model output that contributed to the decision fits that description uncomfortably well. A denial letter citing a generic lack of objective findings, where the file shows a score or a predicted duration that drove the referral, invites the argument that the stated reason was not the operative one. The practical standard to design against: a claimant should be able to read the letter and understand what actually happened to their claim, and the administrative record should not contain a determinative input they were never told about.

How does disability discrimination law reach a claims model?

Through several doors at once, which is why the exposure is larger than a single statute suggests. Benefit administration by an employer or its agent sits near employment discrimination law; insurance sold and administered in a state sits under that state's unfair claims practices and anti-discrimination rules; federally funded health programmes carry their own disability non-discrimination obligations; and the newer state AI statutes attach duties to consequential decisions, which insurance and benefits usually are by name. Add the insurance regulators' governance expectations for AI use by insurers. The common thread across all of them is documented governance and testing rather than a prohibition on modelling, and the organisation that cannot produce testing evidence is the one with a problem in every forum simultaneously.

Our model never sees disability status. Isn't that the point?

In disability claims the target variable is the disability, so the usual protected-attribute-blindness argument does not transfer. The model is explicitly reasoning about a medical condition and its functional consequences, and the risk is not that it uses a forbidden attribute but that it treats some conditions systematically worse than others. The recurring pattern in this line is conditions with limited objective imaging or laboratory confirmation — chronic pain, autoimmune and post-viral conditions, psychiatric conditions, long-tail neurological complaints — where historical files contain more denials, so a model trained on historical outcomes learns to deny them faster. That is a legitimate statistical inference about past behaviour and a serious legal problem: it reproduces a pattern many of these regimes exist to interrupt. Test by condition category and by claimant age and sex within it, and look at outcome distributions, not just accuracy.

Are duration guidelines a model, and do they need the same governance?

Commercially licensed return-to-work duration tables predate machine learning, and many are now sold with predictive layers on top. Treat any of them that influences a decision as an internal guideline the claimant is entitled to ask about, whichever technology produced it. Two failure modes are common. The first is drift from reference to rule: a table published as a median for an uncomplicated case becomes the date on which benefits stop, with the burden silently shifted to the claimant to justify the excess. The second is stacking: the guideline informs the reviewer, the reviewer's note trains the next model, and the model then confirms the guideline. Document which guideline version applied to each claim, and keep the decision to depart from it as an ordinary, unremarkable act rather than an exception that requires escalation.

What should a defensible claim file contain?

Enough that someone reading it two years later can reconstruct the decision without asking the vendor anything. Specifically: which model versions and guideline versions touched the claim and when; what each produced; what the human reviewer did with it, including any override and the reason; the medical evidence relied upon, identified individually rather than as a bundle; and the correspondence sent. Keep the model inputs and outputs in the claim record itself rather than in an analytics system with a shorter retention period than the claim — the most common evidentiary failure in this area is not a bad decision but an unreproducible one, where the file shows a referral with no recoverable cause and the organisation cannot say why the claim was ever pulled for investigation.

AI Disability Benefit Claim Denials 2026: Discrimination and ERISA Exposure

The Decision Nobody Recorded as a Decision

Ask a benefits operation where AI touches a claim and the answer is usually "nowhere that matters — a person decides." Follow the claim instead. It arrives, is scored, and on the strength of that score is routed to a fast-track queue or an investigative one. The investigative queue orders an independent medical examination, a vocational assessment, sometimes surveillance. The fast-track queue approves. Two claimants with similar files receive materially different processes, and the fork was chosen by a model that never produced a denial and never appears in a denial letter.

Every regime that governs this work cares about that fork. Discrimination law cares because scrutiny is a burden distributed unevenly. Claims-handling regulation cares because it prescribes how claims are investigated, not only how they are decided. And a reviewing court asked whether the process was full and fair cares a great deal about a step in the process the claimant was never told existed.

Why the Usual Bias Defence Does Not Work Here

In hiring or lending, teams defend a model by showing it never sees race, sex or disability. In disability claims the model is reasoning about the medical condition directly — that is its job. So the question moves from which attributes the model can see to which conditions it treats worse, and the honest answer usually falls out of the training data without anyone intending it.

Historical claim files are a record of past decisions, not of true functional capacity. Where those files reflect a long institutional scepticism toward conditions without a clean objective marker, a model fitted to them learns the scepticism as signal. It will predict denial for those claims accurately, because the organisation did deny them — and it will be measured as accurate by the very outcomes it is reproducing. This is the central measurement trap of the category: label leakage from the decision you are trying to audit.

Eight Questions a Fiduciary Should Be Able to Answer

Which models and guideline versions touched this claim, on what dates?
What did each output, and what did the reviewer do with it?
What was the stated reason for the adverse determination, and is it the operative one?
Which internal rules or criteria were relied upon, and would we produce them on request?
How do referral, surveillance and denial rates vary by condition category?
How do they vary by claimant age and sex within condition category?
What is the override rate, and is it high enough to call the human review real?
Can we reproduce the score for this claim today from the retained record?

The override rate is the most diagnostic item on that list. A reviewer who accepts the model's recommendation in almost every case is not providing oversight, whatever the procedure document says, and the figure is one query away in any system with an audit trail. Organisations that measure it are frequently surprised, in both directions.

Surveillance Triage Deserves Its Own Review

Of every downstream action a claims model can trigger, surveillance is the most intrusive and the least examined. It is typically governed by an internal referral policy rather than by anything a claimant sees, and the referral criteria are exactly the sort of internal guideline a full-and-fair-review request reaches. When the referral is model-driven, two additional problems attach: the distribution of surveillance across condition categories becomes a measurable disparity, and the model's inputs may include social-media or lifestyle signals whose collection has its own privacy consequences.

Treat a surveillance referral as a consequential decision in its own right — with a documented reason, human authorisation that is not a rubber stamp, and periodic distribution testing — rather than as an investigative detail beneath the level of governance.

Questions Carriers, TPAs and Plan Sponsors Ask

We are the plan sponsor, not the carrier. Does any of this reach us?

It reaches you through the fiduciary duties attached to plan administration and through your selection and monitoring of the entity doing the work. Delegating claims administration does not delegate the duty to select and monitor prudently, and a monitoring process that never asks how claims are adjudicated is thin. The practical ask is small and almost never made: request the vendor's description of any automated tools used in adjudication, the governance around them, and any outcome testing by condition category. A vendor unable to answer is telling you something useful, and the request itself is evidence of monitoring if the question is ever litigated.

Does telling the claimant a model was used invite challenges?

The alternative invites worse ones. A claimant who discovers a determinative model input during discovery gets two arguments — the substance of the decision and the concealment — and the second tends to colour a reviewer's reading of the first. Disclosure also has a quieter benefit: the specific-reason requirement forces the letter to name what actually drove the outcome, which surfaces internally when the letter cannot honestly do so. Organisations that adopt plain-language explanations generally find appeal volume roughly stable and appeal quality higher, because claimants send evidence addressed to the real issue instead of guessing.

How do we test for disparity when we do not hold demographic data?

Start with what you certainly hold: condition category, occupation class, claim duration, referral and surveillance rates, approval and termination rates. Condition-level disparity is the legally salient axis here and needs no protected attribute at all. Age and sex are typically already in the file for underwriting or benefit-calculation reasons. For anything genuinely absent, a separated, access-controlled evaluation dataset used only for bias measurement is the pattern regulators and the newer AI statutes contemplate, and it is not the same thing as adding a demographic feature to the model.

The model is a vendor's and they call it proprietary. Now what?

Proprietary is a commercial position, not a legal exemption from the disclosure and governance duties that sit on you. Negotiate the needed rights before signing: documentation of the model's purpose and inputs, outcome testing by cohort, notification of version changes, retention of per-claim inputs and outputs in your record, and the right to disclose whatever a regulator or a claimant's representative is entitled to receive. Where a vendor refuses, the meaningful question is whether you can defend a decision you cannot explain — and the answer determines whether that product is usable in adjudication at all, however good its accuracy numbers look in the deck.

What is the single highest-value change to make first?

Write the model's contribution into the claim file at the moment it happens. Not a policy, not a committee — a field. Most of the exposure in this area is evidentiary rather than substantive: the decisions are often defensible and the files cannot show it, because the score lived in a pipeline with a ninety-day retention and nobody can now say why a claim was routed for investigation in the first place. A recorded model version, input snapshot, output and reviewer action costs one schema change and converts an unreconstructable decision into an ordinary defensible one.

Run the Override Rate This Week

One query: of the claims where a model recommended referral, denial or a duration, how often did the human reviewer reach a different answer? If the number is close to zero, the human-in-the-loop in your procedure document does not exist in your operation, and every defence built on it is built on the query you have not run.

Then break the outcome rates down by condition category. Those two numbers decide how much of the rest of this list you actually need.